v0.2.0 — Embedded RAG (no servers), exact API lookup, audit remediation

Highlights

The RAG stack is now fully embedded — no Docker, no Redis, no Ollama. clone → uv sync → run, with prebuilt indexes attached to this release.

• lookup_api (new tool) — exact, lossless answers for API questions (enum values, endpoints, schema fields) from a SQLite FTS5 index over 213 parsed OpenAPI specs — including the Central NAC Service spec (cnac-mac-reg / visitor / named-mpsk / dpp)
• search_docs rebuilt — hybrid (vector + BM25, RRF-fused) retrieval via LanceDB + in-process fastembed (nomic-embed-text-v1.5 with proper task prefixes), over a full 53,052-chunk corpus (the old index was silently missing 3 sources)
• find_tool — tool catalog search now embedded too (tools.lance)
• Safety fixes from the 2026-06-03 audit — destructive ops (reboot_device, poe_bounce, port_bounce, disconnect_client) now elicit confirmation correctly and fail closed; router dispatch forwards Context so confirmations work through invoke_tool

Measured retrieval quality (bundled eval, tests/eval/)

Metric	v0.1.x (Redis, vector-only)	v0.2.0 (embedded hybrid)
api_exact	0.50	1.00
howto_recall@5	0.80	0.90
source_hit@5	0.50	0.90
mrr	0.34	0.90

Installing the prebuilt index

curl -L -o /tmp/rag-index.tar.gz \
  https://github.com/secure-ssid/centralmcp/releases/download/v0.2.0/centralmcp-rag-index-v0.2.0.tar.gz
tar -xzf /tmp/rag-index.tar.gz -C /path/to/centralmcp   # extracts data/

Or rebuild locally: uv run python ingestion/ingest_docs.py && uv run python scripts/ingest_tools.py (embeds 53k chunks in-process — several hours on CPU). See data/INDEX-MANIFEST.json inside the tarball for provenance.

Upgrading from v0.1.x

• The default RAG backend is now lancedb. Redis Stack + Ollama remain supported: set CENTRALMCP_RAG_BACKEND=redis (see the README's optional server backend appendix).
• qdrant-client dependency removed.
• New env knobs: CENTRALMCP_RAG_BACKEND, CENTRALMCP_EMBED_PROVIDERS.

🤖 Generated with Claude Code